Download A New Method to Detect Related Function Among Proteins

doi:10.1016/S0022-2836(02)00811-2 available online at http://www.idealibrary.com on
w
B
J. Mol. Biol. (2002) 323, 387–406
A New Method to Detect Related Function Among
Proteins Independent of Sequence and Fold Homology
Stefan Schmitt1,2, Daniel Kuhn1 and Gerhard Klebe1*
1
Inst. of Pharmaceutical
Chemistry, Univ. of Marburg
Marbacher Weg 6, D-35032
Marburg, Germany
2
Structural Chemistry
Laboratory, AstraZeneca
R&D Mölndal, S-43183
Mölndal, Sweden
A new method has been developed to detect functional relationships
among proteins independent of a given sequence or fold homology. It is
based on the idea that protein function is intimately related to the recognition and subsequent response to the binding of a substrate or an
endogenous ligand in a well-characterized binding pocket. Thus, recognition of similar ligands, supposedly linked to similar function, requires
conserved recognition features exposed in terms of common physicochemical interaction properties via the functional groups of the residues
flanking a particular binding cavity. Following a technique commonly
used in the comparison of small molecule ligands, generic pseudocenters
coding for possible interaction properties were assigned for a large sample
set of cavities extracted from the entire PDB and stored in the database
Cavbase. Using a particular query cavity a series of related cavities of
decreasing similarity is detected based on a clique detection algorithm.
The detected similarity is ranked according to property-based surface
patches shared in common by the different clique solutions. The approach
either retrieves protein cavities accommodating the same (e.g. co-factors)
or closely related ligands or it extracts proteins exhibiting similar function
in terms of a related catalytic mechanism. Finally the new method has
strong potential to suggest alternative molecular skeletons in de novo
design. The retrieval of molecular building blocks accommodated in a
particular sub-pocket that shares similarity with the pocket in a protein
studied by drug design can inspire the discovery of novel ligands.
q 2002 Elsevier Science Ltd. All rights reserved
*Corresponding author
Keywords: functional comparison of proteins; cavity comparison; data
mining; de novo design; physicochemical properties
Introduction
Genomic sciences provide us with the sequences
of the entire human genome and other important
microbial pathogens.1 – 4 By means of proteomics
and powerful bioinformatic tools5 – 9 it is hoped to
determine gene variants that contribute to various
multifactorial diseases or to detect genes that exist
in certain infectious agents but not in humans. As
a consequence, a large number of new suitable
targets for drug intervention may be discovered.
Presently structural genomics embark on highthroughput X-ray crystallography and NMR spectroscopy to obtain a comprehensive view on the
world of protein structures.10 This tremendous
increase in experimentally resolved protein structures will be accomplished by an even larger
number of protein structure models computed by
E-mail address of the corresponding author:
[email protected]
homology modeling.11,12 The challenge, once this
entire body of structural knowledge has been produced, is to extract relevant information about the
properties and the functional role of individual
proteins detected in particular organisms. This
new strategy that seeks for the spatial structure of
a protein prior to the knowledge of its actual function might provide the challenging opportunity to
identify new proteins as potential drug targets.
These developments call upon methods to infer
protein function directly from 3D structure. The
geometry of a protein usually carries information
about its biochemical function on a molecular
level, e.g. as a serine protease or an oligonucleotide
binding protein. However, its influence on the biological function of a cell or even more on an entire
organism can only be determined in a comprehensive study where a complete series of different
experimental evidences are brought together.
Nevertheless, even to infer function in its biochemical sense is not straightforward, since protein
0022-2836/02/$ - see front matter q 2002 Elsevier Science Ltd. All rights reserved
388
function is not necessarily confined to a particular
fold and it is often enough not apparent at the
sequence level.13 Protein function, in particular of
enzymes, is often intimately connected with the
recognition and chemical modification of endogenous ligands such as agonists, antagonists, effectors
or substrates. This recognition usually occurs in
well-characterized clefts or cavities of protein surfaces. It has been shown for enzyme active sites
that over 70% of these sites can be easily detected
as the largest cleft on the surface.14 For the latter
class of proteins, elementary steps of a chemical
reaction are proceeded that require a strictly
defined spatial arrangement of molecular recognition determinants in the enzyme active site to
accommodate and spatially arrest the substrates.
Very similar requirements exist for the specific recognition of co-factors in proteins or the binding of
endogenous ligands (e.g. the biogenic amines) in
signal transduction cascades.
This idea that molecular recognition patterns
may be conserved throughout the binding pockets
of proteins of similar function stimulated us to
develop a new method to detect relationships
among proteins. The approach involves the automatic detection and extraction of putative binding
sites from proteins. Subsequently, the actual constitution of these extracted sites has to be determined
and translated into molecular descriptors that are
not simply based on atomic coordinates of the
binding-site exposed residues but on associated
physicochemical properties. Finally, the thus attributed descriptors serve as a base for the mutual
comparison of different binding sites. All aspects
of the algorithmic development have been conceived with regard to an efficient handling of the
huge data samples of protein structures. Thus, all
steps of retrieval, reduction and analysis of raw
data have been drafted in a way to operate automatically, avoiding manual interference. The
screening of a particular binding site against a
database of several thousand binding cavities
allows retrieval of proteins of similar function
together with possibly bound ligands. In turn, the
thus achieved indirect retrieval of bound ligands
or ligand portions accommodated in structurally
related binding cavities or subcavities might reveal
interesting suggestions on putative bioisosteric
fragments of ligands. Such ideas are extremely
valuable in structure-based de novo design of
novel leads.
Several approaches have been described in the
literature to detect structural and/or functional
relationships among proteins. Such similarity can
be classified on three levels. The earliest comparative algorithms are based on sequence
information,15 – 17 such as FASTA or routines
implemented into databases (such as OWL or
SWISS-PROT).18 – 21 While high structural homology
usually matches with pronounced sequence homology, the reverse that low sequence homology parallels with structural dissimilarity, is not
necessarily given. Accordingly more recent
Detecting Related Function Among Proteins
methods determine protein similarity in terms of
the overall 3D-fold. Proceeding from sequence
similarity searches to comparisons based on spatial
coordinates requires more complex algorithms to
encode protein structural information. Considering
two proteins as rigid objects, a translation and
rotation matrix has to be found for spatial superimposition. This can be computed purely in geometrical terms, however more reliable solutions are
obtained if the coordinates are associated with
some predefined properties. Algorithms considering all atomic positions are far too demanding for
fast similarity searches. Therefore approximative
representations are used, e.g. based on Ca-atom
coordinates. The scope of similarity search algorithms following these concepts ranges from distance matrix methods,22,23 complete common
subgraph searches24 and geometric hashing
techniques25 to genetic algorithms.26 – 28 Some
enhanced techniques include precalculated properties in the assignment.29 – 34
In most of the above-mentioned techniques, the
protein structures are purely described in terms of
Ca-atom coordinates and geometric similarities are
computed with respect to distance and angular
relationships, occasionally complemented by
sequence, secondary structure or amino acid property information. With respect to the recognition
of a ligand in its binding site, this reduction to Ca
coordinates appears rather crude and limiting.
However, considering all protein atoms is computationally hardly tractable. Thus, a third level of
similarity search techniques has been developed
that focuses on smaller subregions. They provide
a compromise between computational tractability
and structural complexity. The programs TESS35
and ASSAM36 use geometric hashing or clique
detection, respectively, to retrieve templates of
pre-defined 3D amino acid patterns. Through this
predefinition of a particular pattern, the
approaches are to some extent biased with respect
to that what we can expect to be retrieved. A combination of sequence comparison to detect highly
conserved residues along with recursive distance
matrix alignments of coordinate sets representing
“centers of functional activities” retrieves a large
number of common substructures in proteins37
independent from a particularly selected input
structure. While most of these methods search the
entire protein structure for common motifs, recent
developments target the binding sites only to
identify similarities that could support and assist
drug design. They require assumptions about the
spatial location and mutual superposition of such
binding sites. In some studies, this initial step is
performed manually38 or based on commonly
bound ligands or co-factors.39 Recently Stahl et al.40
reported on the analysis of 176 preselected zinc
metalloproteinases that have been clustered in
terms of solvent-accessible surface patches
assigned to different physicochemical properties
using a self-organizing neuronal net. The zinc
active sites could be discriminated from other
Detecting Related Function Among Proteins
surface depressions found on these enzymes.
Approaches operating independent of bound
ligands have been described by Fischer et al.41 and
Rosen et al.42 They use geometric hashing for similarity searching, but different parameters are used
to describe the binding site. Fischer et al. use SURFNET spheres43 to generate a negative image of the
active site and use coordinates of this representation for searching and comparison. Rosen et al.
first proposed the use of generic coordinates to
describe putative ligand binding. They are defined
as sparse critical points,44 assigned to patches of
the Connolly surface considering its local curvature. Although this approach is less accurate compared to similarity searches based on discrete
atomic coordinates and performance is reduced
once the critical points are assigned to properties,
their ideas inspired us to follow a related concept
using a reduced set of assigned coordinates, however combined with strategies usually applied in
the field of similarity searches among small molecule ligands.45,46
In the present paper, we describe the algorithmic
development of a new concept to compare binding
pockets in proteins. A new object-oriented database Cavbase, fully integrated with the receptorligand database Relibase47,48 has been developed.
As input, the method uses a large data set of
cavities. For this purpose, any suitable algorithm
described in the literature can be used to detect
and extract surface depressions. Subsequently,
descriptors to encode the molecular recognition
determinants of a binding site are assigned. A clique detection algorithm is used to compare these
binding site descriptions, together with a sophisticated ranking of the obtained multiple solutions.
Finally, a representative set of example problems
is used to assess the scope and demonstrate the
power of the present method in particular with
respect to data mining.
Theory and Algorithms
Cavity extraction and descriptors for
cavity properties
In the present study, the comparison of a large
binding site sample is attempted. To achieve this
objective, special requirements with respect to the
definition of binding-site regions, the assignment
of reliable descriptors and the subsequent processing of the retrieved information has to be met. A
number of programs have been developed to locate
depressions on protein surfaces as putative binding sites. They apply different algorithmic concepts, such as flood filling techniques,43,49 gridbased50 – 52 or alpha shape-based approaches.53 – 56
Usually these programs operate on raw PDB data
and produce either graphical output or new flatfile information. Facing the automatically extracted
binding sites with areas known to bind a ligand
reveals convincing agreement and underlines the
389
reliability of these programs. However, to avoid
significant pre- and post-processing we decided to
access directly the pre-processed data stored in
the object-oriented database Relibase.57,58 We
implemented the cavity detection algorithm of
Ligsite51 into Relibase. Additional information
from Relibase was used to attribute appropriate
cavity descriptors. The accordingly extracted information has been deposited with the new database
module Cavbase, sharing similar architecture with
Relibase. It has been equipped with a graphical
interface for data evaluation. In the Ligsite algorithm, the protein under consideration is embedded
into a regularly-spaced Cartesian grid (0.5 Å grid
spacing). Any grid points, represented by 1.5 Å
probe spheres, penetrating into protein atoms
within their van der Waals radius are discarded as
solvent-inaccessible. In order to determine which
solvent-accessible grid points fall into a cavity, Ligsite scans along the three Cartesian axes and the
four cubic diagonals for regions, that terminate
the scan directions on either ends by protein
atoms. A counter is set to the number of scan directions terminated by protein atoms. This counter is
used as a measure for the burial of the solventaccessible grid points. It spans a range from 0
(fully solvent exposed) to 7 (deeply buried).
A protein must have a least one cluster of adjacent grid points comprising more than 320 grid
points (approximately 40 Å3) with a degree of burial ^ 4. If no cluster of this size could be detected,
we reduced the degree of burial for grid points to
be considered in the cluster to values of three or
two subsequently. This allows us to detect more
shallow cavities. If present, neighboring grid
points are merged into such starting clusters. The
size threshold of a thus obtained cluster to be
accepted has to be greater than 40 Å3. This allows
it to accommodate at least one water molecule.
All surface-contacting grid points of a cluster,
apart from those oriented towards the solvent, are
used to approximate the cavity surface. If one
atom of an amino acid residue falls closer than
1.1 Å to a protein surface-contacting grid point,
the amino acid is classified as a cavity-flanking
residue. These data are used to represent the basic
geometric shape of cavities in the database.
To compute spatial similarities of cavities across
a large sample set an algorithm based on a
restricted number of input coordinates is required.
Thus, considering all coordinates of every cavityflanking residue would be intractable. For the
same reason, most approaches in the literature
operate on reduced spatial information, e.g.
Ca coordinates instead of entire residues (see
above). This strategy completely neglects the type
of interactions a particular residue could possibly
perform to an accommodated ligand. As a consequence, we decided to condense the physicochemical properties of the cavity-flanking residues into
a restricted set of generic pseudocenters corresponding to five properties essential for molecular
recognition: hydrogen-bond donor (DO), acceptor
390
Detecting Related Function Among Proteins
Table 1. Encoding of the rules for the pseudocenter assignment
Side-chain
Amino acid
Pseudocenter (type)
Origin atoms
Ala
Aliphatic
CB
Aliphatic
Donor
CB, CG, CD
NE
Donor
Donor
NH1
NH2
Acceptor
OD1
Donor
ND2
Acceptor
OD1
Acceptor
OD2
Aliphatic
CB, SG
Acceptor
OE1
Donor
NE2
Acceptor
OE1
Acceptor
OE2
His
PI
DON_ACC
DON_ACC
CG, ND1, CD2, CE1, NE2
NE1
NE2
Ile
Aliphatic
CB, CG1, CG2, CD1
Leu
Aliphatic
CB, CG, CD1, CD2
Aliphatic
CB, CG, CD, CE
Donor
NZ
Aliphatic
CB, CG, SD, CE
Arg
Asn
Asp
a
Cys
Gln
Glu
Lys
Met
(continued)
391
Detecting Related Function Among Proteins
Table 1 Continued
Side-chain
Amino acid
Pseudocenter (type)
Origin atoms
Phe
PI
CG, CD1, CD2, CE1, CE2, CZ
Pro
Aliphatic
CB, CG, CD
Ser
DON_ACC
OG
Aliphatic
CD2
DON_ACC
OD1
PI
CG, CD1, CD2,
NE1, CE2, CE3,
Donor
CZ1, CZ3, CH
NE1
Thr
Trp
PI
DON_ACC
CB, CD1, CD2,
CE1, CE1, CZ
OH
Val
Aliphatic
CB, CG1, CG2
Pep
Acceptor
Donor
PI
O
N
C
Tyr
b
Particular atoms or functional groups define the coordinates for the different pseudocenters. Arrows indicate the origin and directionality for the mean H-bonding property exposure (v, see Figure 1) in the case of donor, acceptor and mixed donor/acceptor pseudocenters. Triangles and rings show the location of aliphatic and pi centers, respectively. The latter centers expose their properties
above and below a best plane through the atoms of corresponding functional group. Aliphatic centers can be shifted towards atoms
that are more exposed to the molecular surface of the cavity (see the text).
a
Thiol or disulfide bridge.
b
Peptide bond.
(AC), mixed donor/acceptor (DA, e.g. hydroxyl
groups or side-chain nitrogen atoms in histidine),
hydrophobic aliphatic (AL) and aromatic (PI) contact. This crucial assignment of pseudocenters to
the individual amino acids obeys the rules summarized in Table 1.
The phenyl groups in Phe and Tyr are described
by one PI center, respectively, representing the center-of-mass of the six ring carbon atoms. Similarly,
PI centers are generated using all ring atoms in
the side-chain of His and Trp, respectively. The
oxygen atoms of hydroxyl groups in Ser, Thr and
392
Detecting Related Function Among Proteins
Table 2. Cut-off values for the angles between v and r
(see Figure 1)
Pseudocenter (type)
Cut-off (8)
Donor
Acceptor
Donor/Acceptor
PI
100
100
120
60
Pseudocenters with higher values are discarded from the set
of pseudocenters that define the cavity.
Figure 1. Two vectors, v and r, define the exposure of
a particular physicochemical property. The standard
vector v represents the mean direction of its property
exposure, which matches in the case of an AC center,
assigned to the carbonyl oxygen, the vector along the
projected CvO axis. r is derived as the normalized summation vector of all vectors oriented from the oxygen to
all neighboring surface grid points (S ) within a predefined distance (dmax). The angle between v and r
(Table 2) is taken as a criterion whether a pseudocenter
exposes its property to a putative ligand in the cavity or
whether it is removed from the list of cavity defining
pseudocenters. In this example, the DO center for the
backbone nitrogen would be removed.
Tyr can act as hydrogen-bond donor, and via their
lone pairs as acceptor. The atomic coordinates of
hydroxyl oxygen atoms are therefore assigned to
the mixed donor/acceptor property (DA type). A
similar situation holds for nitrogen atoms in His
residues. The assignment of protonation states is
difficult based on X-ray data in particular if the
pKa of a functional group is about 6.5 as for His.
Again, the assignment of DA pseudocenters to the
atomic coordinates of the two nitrogen atoms is
anticipated as best compromise. Furthermore, the
protonation states of the carboxy groups in Glu
and Asp side-chains are sometimes difficult to
define. However, on a first glance in our approximate model, these oxygen atoms are assumed to
display AC centers. Peptide bonds are represented
by three different types of centers, AC, DO, and PI
for carbonyl oxygen, nitrogen, and carbonyl carbon, respectively. PI centers are assigned above
and below a local best plane through the atoms of
the peptide bond. A similar assignment would
also be justified for the atoms of a terminal carboxy,
carboxamide, and guanidino group in the sidechains of Asp, Glu, Asn, Gln, and Arg, however
the present version neglects such assignments of
PI centers. As further approximation we neglect
presently hydrogen-bonding properties of sulfur
atoms in Cys and Met. They are described similarly to aliphatic carbon atoms. Aliphatic centers
(AL) are attributed to the side-chains of Ala, Arg,
Cys, Ile, Leu, Lys, Met, Pro, and Val according to
the centers-of-mass formed by their aliphatic carbon (and sulfur) atoms. According to this pro-
cedure the assigned coordinates of AL centers can
vary significantly with the side-chain length and
adopted conformation. This could possibly result
in unreasonable spatial positions of the pseudocenters and therefore bias the contribution of the aliphatic properties in an unreasonable manner. To
focus more strongly on the contributing part of surface-exposed aliphatic side-chains, only those carbon (and sulfur) atoms are considered in the
calculation of the geometric mean that expose
their property onto a surface area greater than 1 Å
(approximately five grid points) within a distance
of 3.5 Å. This reflects the scope of aliphatic interactions. The contribution of each considered atom
to the geometric mean is finally weighted according to its distance from the closest surface-contacting grid points. As a result, AL centers are shifted
in direction towards those aliphatic side-chain
atoms that are placed next to the cavity surface. In
this area they contribute most to the exposed aliphatic property. Following these rules all atoms of
the cavity-flanking residues are converted into generic pseudocenters. They express the features of the
20 different amino acids in terms of five wellplaced physicochemical properties.
Subsequently, the assigned pseudocenters are
examined with respect to their surface exposure.
This step tries to verify whether a particular interaction property could possibly form an interaction
to a bound ligand. To assess their favorable
exposure, the angle between the following two vectors v and r, assigned to each pseudocenter, is
computed.
The first vector v describes the mean orientation
along which a particular interaction could be
formed. To retrieve information about given orientational preferences, data stored in the IsoStar
database59 have been consulted in detail. For
example, for a DO pseudocenter, generated at the
position of a nitrogen connecting two carbons, the
vector v orients along the assumed NH bond. For
PI centers two vectors v are generated perpendicular to the plane defined by the atoms contributing
to the PI center. The pseudocenter attributed to
the position of a terminal oxygen acceptor AC is
represented by a vector v oriented along the projected C-O bond axis. Next, a second vector r is
computed as normalized summation vector of all
vectors that point from a particular pseudocenter
to all neighboring surface-contacting grid points
that fall into a 3 Å sphere around this center. This
393
Detecting Related Function Among Proteins
Figure 2. The shape and the
properties of a binding site are
determined by the amino acids (I)
flanking the site. These amino acids
are translated into a set of pseudocenters (III), displayed as colored
spheres.
Every
pseudocenter
exposes its property onto a certain
surface patch (III). The following
color scheme is used: H-bond
donor (blue), H-bond acceptor
(red), ambivalent donor/acceptor (green), hydrophobic aliphatic (white) or aromatic (orange). Only those pseudocenters are considered that expose their physicochemical properties onto the surface (II).
roughly describes how the cavity surface nestles
against a particular pseudo center. The angle
enclosed by the two vectors serves as a criterion
whether a particular pseudocenter is considered
in the analysis or discarded (Figure 1). Table 2
summarizes the cut-off values for these angles.
They were calibrated according to populated areas
found in IsoStar.59 AL centers are not examined
with respect to these directionality criteria, mainly
because they are assumed to interact isotropically
in space through van der Waals forces. Finally, all
surface-contacting grid points describing the cavity
surface are attributed to adjacent pseudocenters. In
this step, at first a distance criterion of # 3 Å has to
be matched and second the above-described angular selection criterion must be met. Once assigned,
the various surface-contacting grid points are
ascribed to one of the five physicochemical properties represented by the adjacent pseudocenter
(Figure 2). According to this procedure the properties of all pseudocenters are projected onto the cavity surface. In case that surface points fall next to
more then one pseudocenter within 3 Å, assignment to the closest center is accomplished. As
final result, each Ligsite-extracted surface
depression is represented by a set of residue-attributed pseudocenters. The cavity surface, approximated by the set of surface-contacting grid points
is decomposed into surface patches assigned to
one of the five physicochemical properties exhibited by the most adjacent pseudocenter. This
abstracted description represents the input data
for the cavity comparison.
Similarity searching algorithm
The detection of a common motif in two cavities,
represented by the above-defined descriptors, corresponds to the problem of finding a complete
common subgraph in two sets of descriptors. Solutions to this problem are discovered by clique
detection algorithms.60 – 62
A 3D arrangement of pseudocenters can be
regarded as a graph C for which the nodes (c [ C)
correspond to pseudocenters and the edges correspond to distances between two pseudocenters
(d(ci; cj), with i,j ¼ 1,..., lCl). Given a pair of graphs
A and B, i.e. by nodes (a [ A and b [ B) and
edges (d(ai; aj) and d(bk; bl)) that describe two cavities A and B, a new graph G can be defined
according to the following protocol: (1) construct
pairs between nodes taken from A and B, in such
a way that the nodes (ai and bk) correspond to the
same property, i.e. allowed combinations are DOiDOk, ACi-ACk, etc. Pseudocenters assigned to the
mixed property DA can also form pairs with DO
Figure 3. Schematic representation to explain the matching and
scoring scheme of two binding
pockets. Cavity B shares two contiguous subsets (BI and BII) of pseudocenters in common with a subset in
cavity A. Either the types or distances among the pseudocenters
are similar. Therefore both subsets
will be detected by the clique algorithm, which is solely based on distance and property information. To
determine, which substructure
pattern match possesses physicochemical relevance, the corresponding pseudocenters are superimposed and a score is calculated
which takes the mutual overlap of
the binding site surface patches
into account.
394
Detecting Related Function Among Proteins
Figure 4. A cavity of interest
(query cavity) is compared against
a database of probe cavities. For
each mutual comparison 100 clique
solutions are generated and scored
according to the overlapping surface patches. The best-scored clique
solution of each individual comparison is kept and finally all best –
scored solutions are sorted to detect
the best score among all generated
comparisons.
and AC centers, thus additional combinations such
as DOi-DAk, ACi-DAk, vice versa and DAi-DAk are
allowed. The new nodes gi,k in G correspond to all
allowed pair combinations (ai; bk). (2) Edges in G
are defined as pairs of nodes (gi,k; gj,l) for which
the actual distances among the generic pseudocenters match within a predefined tolerance, i.e. gi,k
and gj,l are connected if d(ai; aj) < d(bk; bl). The distance tolerance has been set to 2 Å to cope for
spatial uncertainties in the pseudocenter positions.
This value has been rationalized by comparing
multiple PDB entries of the same protein however
bound to different ligands. It has to be remembered that such deviations originate either from
the limited accuracy of protein crystal structure
determinations and, even more pronounced, from
conformational differences in the cavity-flanking
residues among related proteins. In addition, only
intercenter distances up to 12 Å have been
regarded. Rationale behind this cut-off is the limitation of our similarity search to distances in the
short and medium range, in particular since the
described uncertainties are likely to increase at
longer distances.
The Bron-Kerbosh algorithm60 has been applied
to find the maximal common subgraph in
G. Reflected back onto the pseudocenters, such a
common subgraph represents a similar spatial
arrangement of properties in two cavities, thus
defining a similar motif. Clique detection algorithms are computationally demanding, since they
scale with N2 for every additional node gi,k. However, the above-described assignments produce a
limited set of descriptors being a satisfactory compromise between required accuracy (number of
centers) and computational tractability (pair-wise
comparison of about 70 centers takes approximately three seconds on a state-of-the-art Linux
processor). Solely considering the distance matrix
among pseudocenters can still produce chemically
unreasonable solutions. E.g. the result from a clique detection is geometrically still reasonable if
equivalent centers from a concave area in one cavity match upon a convex one in a second (Figure 3).
In such cases, the actual superposition reveals a
chemically unreasonable match. Obviously, the
direction of property exposure matters in the comparison. Attempts to consider directionality in the
definition of nodes gi,k, by means of the vectors v
and/or r, did not improve the detection of correct
solutions. Thus, we decided to compute an independent scoring to rank the generated clique solutions. It considers the mutual matching of
assigned surface patches of the two cavities
aligned, according to the shared pseudocenters
detected by the clique algorithm. The following
protocol is accomplished (Figure 4): For each pairwise comparison of cavities, the 100 largest common subgraphs are evaluated. This results in 100
common spatial arrangements of pair-wise matching pseudocenters (individual clique solutions).†
Each clique solution generates a matrix that transforms the matching pseudocenters together with
the associated surface patches onto those of the
reference as best spatial superposition. Subsequently, each generated superposition is analyzed and the overlap in surface points (p [ P),
assigned to the same physicochemical property, is
determined. These surface points originated from
the embedded grid of 0.5 Å spacing, thus the
mutual overlap SAB of points from the two matching surfaces is calculated as:
SAB ¼
X
sv
ðfor all sv $ 0:7Þ
v
s¼
rai þ rbk
lPai l þ lPbk l
with
rai ¼ l{pai ldðpai ; pbk Þ # 1:0}l
† A value of 100 is the best empirically determined
compromise between computational effort and achieved
coverage of solutions.
395
Detecting Related Function Among Proteins
Figure 5. Schematic drawing of two possibilities to
superimpose minor surface patches P generated by the
Acceptor centers ai and bk of two carbonyl groups. Both
superpositions result in very similar surface matches
and will therefore contribute equally to SAB. Thus R1 cannot discriminate between both solutions, R2, however,
will favor solution II due to an implicit consideration of
the directionality via the RMSD value.
and
rbk ¼ l{pbk ldðpbk ; pai Þ # 1:0}l
The degree of mutual overlap in surface patches
is expressed by the number of surface points that
fall next to each other below a distance threshold
of 1.0 Å. To avoid consideration of strongly fragmented surface patches, the mutual overlap of
patches is only counted if at least 70% of the
matched surface patches†, corresponding to a
pseudocenter pair, fall next to each other below
1 Å. Accordingly, if n common pseudocenter pairs
have been detected in the subgraph analysis, SAB
exceeds maximally to a value of n and minimally
to a value of SAB ¼ 0.7n. Those of the pseudocenter
pairs that pass the overlap criterion of adjacent surface patches ($ 70%) define a subset of equivalent
centers representing the physicochemical properties shared by both cavities. In a subsequent refinement step, a new transformation matrix is
computed, however, only considering those pseudocenters that passed the above-defined overlap
criterion. Finally, a new ranking is calculated by
determining the matched surface points approaching each other below 1 Å after the second transformation has been performed.
This procedure is followed for the above-mentioned 100 best clique solutions in the pairwise cavity comparison. The procedure reveals 100
improved solutions with pairs of equivalent pseudocenters. Out of these, the solution with the highest SAB value is stored together with the
corresponding set of n equivalent pseudocenter
pairs. On the mean, the refinement and scoring
procedure requires additional 100 seconds for two
medium sized cavities (ca. 800 Å3) on a state-ofthe-art Linux processor.
For each cavity in a test set such a pairwise comparison with a query cavity is performed. The various solutions, obtained for the entire sample of
† This value is an empirical estimate that does not
penalize conformational deviations too strongly, but
considers similar property exposure onto the surface.
test cavities, are ranked according to SAB and n.
The query cavity to be compared with the test
sample could comprise all pseudocenters representing the entire binding site or could be
reduced to a pseudocenter subset, e.g. to a specifically edited sub-pocket. Two figures-of-merit (R1
and R2 ) are considered to rank the entire set of cavities with respect to their similarity with the query
cavity:
R1 ¼ SAB
and
R2 ¼
SAB 2 0:7n
RMSD
where RMSD corresponds to the root mean square
deviation of the matched pseudocenter pairs used
for superpositioning. Using R1, the list of pairwise
comparisons is simply sorted in terms of the size
of their achieved surface overlap, accordingly cavities that share multiple surface patches in common
with the query cavity will occur at the top of the
list. Visual inspection of the top-ranked cavity
matches disclose some deficiencies while focusing
entirely on R1. In particular fragmented and disconnected motifs of rather small surface patches
produce a mutual overlap that is not very conclusive with respect to shared property distributions.
Figure 5 illustrates such a situation where two
possible superpositions with similar s and accordingly SAB result in equivalent R1 values, even so
the match of pseudocenters is quite unsatisfactory
(Figure 5, right). Obviously, the pure summation
over common surface patches has to be weighted
by the total number of contributing pseudocenters
and their spatial matching accuracy. Thus, we
rank their mutual match with respect to the spatial
deviation (RMSD ) achieved in the cavity superpositioning step. The term SAB 2 0.7n describes the
relative size of the overlapping surface patches,
since SAB can adopt maximally a value of n. Any
considerations to include vectors v and/or r as
directionality terms in the scoring and ranking did
not improve the figures-of-merit. In practice, we
sort our comparisons according to R1 and during
visual inspection we consult on purpose R2, SAB
and n. Therefore Cavbase has been equipped similarly to Relibase, with a visualization tool based
on RASMOL.63 It allows one to display and browse
through the top-ranked hits of the similarity analysis in short time.
Prefiltering of data sample
The Ligsite approach used in this study automatically extracts depression on the protein surface
as putative binding site. However, not necessarily
all of them are relevant on a first glance. This
could possibly intricate a critical assessment of the
performance of the new method. Accordingly, for
our initial validation we considered only cavities
that accommodate at least one small molecule
396
Detecting Related Function Among Proteins
Figure 6. Folding pattern and
amino acids comprising the binding
sites of chorismate mutases 1ecm
(E. coli (I)) and 4csm (S. cerevisiae (II))
are shown together with the bound
ligand. In the first case, the binding
site is composed of two different
polypeptide chains, whereas in the
latter case the amino acids originate
from one contiguously connected
chain.
with six up to 50 non-hydrogen atoms, thus covering the range typically found for small drug
molecules.
A further complication arises from redundancies
present in the stored PDB data. Multiple cavity
entries will occur in Cavbase originating from distinct PDB entries, however extracted from the
same underlying protein. In a comparative analysis
these cavity entries will produce a trivial high similarity score. A possible filter to eliminate such
redundancies in the considered cavity data set
could be a pre-selection of PDB entries that are
confirmed as highly diverse with respect to their
overall 3D structure, e.g. similar to the selection
performed by Fischer et al.64 However, we
refrained from pre-selecting PDB entries on the
base of folding patterns, since we wanted to avoid
any selection based on overall protein information,
e.g. folding patterns. Furthermore, the consideration of high sequence similarity as pre-filter is
also not fully reliable since the cavity flanking residues not necessarily originate from one
contiguously connected peptide chain (Figure 6).
Often enough several domains contribute to a cavity and a decision would have to be taken which
peptide chain(s) to consider for sequence comparison. However, comparisons on the sequence level
would not conflict with the precondition that our
approach should neglect any structural information apart from the spatial composition of physicochemical properties exposed to a binding site.
Therefore, the overall cavity data set is divided
into clusters with expected trivial internal similarities pursuing the following protocol:
(1) Each cavity is analyzed with respect to the
peptide chains that contribute cavity flanking
residues.
(2) For each thus detected peptide chain, its
corresponding representative chain is retrieved
from the PDB-select database.65,66 In this database, to each peptide chain found in the PDB, a
representative chain has been assigned according to an all-against-all sequence comparison.
Thus, for a particular cavity under investigation,
a set of one or more representative chains is
attributed resulting in a set of parent sequences.
In case that several chains contribute cavity
flanking residues and all share the same representative chain in common in the PDB-select
database, multiple assignments with the parent
set can occur.
(3) Cavities assigned to the same parent set
according to (1) and (2) are then clustered
together. Thus, the procedure groups cavities
together that are composed of peptide chains
with high sequence similarity. Cavities falling
into the same cluster will likely possess a trivial
similarity score. To avoid such trivial comparisons in our study, we only use one representative cavity from each such formed cluster as
query cavity. Subsequently, it is compared to all
other cavities in the remaining clusters.
Results
Results from the pre-filtering process
The automatic extraction of binding pockets has
been applied to the June 2000 version of the PDB
containing 11,983 entries. Of these, 8627 with a
resolution of 3 Å or better have been considered
after discarding all model-built structures, NMR
and superseded entries or data containing only
Ca coordinates.
A sample set of 31,441 surface depressions
could be detected using our implementation of the
Ligsite algorithm. These were stored in the new
object-oriented Relibase module Cavbase.
For the validation of our method, we considered
only PDB entries that contain at least one ligand
with 6 up to 50 non-hydrogen atoms. The definition of a “ligand” thereby obeys the rules set in
Relibase. Accordingly, 4332 PDB entries remained,
corresponding to 18,402 cavities. This data sample
was further analyzed to consider only cavities that
actually accommodate a ligand. With the precondition that at least one ligand atom must be buried,
only 5448 cavities remained originating from 3626
independent PDB entries. In 696 cases, ligands
(mainly sugars) coincide with flat surface regions
397
Detecting Related Function Among Proteins
Figure 7. The binding site sketch
illustrates that the algorithm can
handle equivalences originating
from different chains. While in the
case of 4csm (II) the cavity is
entirely built-up from a single
chain (B), the active site interactions
in 1ecm (I) reside from two different protein subunits (A and B).
(III) shows the superposition of the
involved amino acids (IIIa) and the
surface patches (IIIb). The example
illustrates the importance to use
generic descriptors, since equivalent H-bonding and hydrophobic
properties are not necessarily
experienced by one particular type
of amino acid.
that are not detected as cavities or the ligand could
be classified as solvent molecule (e.g. benzene)
with more than five atoms. A statistical evaluation
of the volumes found in the data sample of the
18,402 extracted cavities shows that most cavities
fall into a range between 300 and 800 Å3. Regarding only the subset of cavities actually occupied
by a ligand reveals a distribution between 300 and
1500 Å3 with a significant shift of the mean
towards larger volumes (mean about 800 Å3).
Obviously, many of the small pockets (, 500 Å3)
remain unoccupied.
One might expect that cavities retrieved from distinct PDB entries, however originating from the
same underlying protein, possess very similar
shape and thus property distributions. Due to conformational flexibility of proteins, cavities of quite
deviating size can be extracted. To some extent this
is also a result of the parameter settings applied in
Ligsite. Small conformational changes of the protein
found in different PDB entries can trigger the
detection of subpockets or even extended channels
in one entry that is inaccessible in the other. This
results in significant size and shape differences of
cavities although originating from the same parent
protein. The clique detection algorithm used in our
approach for the comparative analysis is capable of
coping with such size differences since it seeks for
Table 3. Equivalent pseudocenter pairs and the involved amino acids of the chorismate mutases structures 1ecm and
4csm used in the cavity matching algorithm.
Corresponding amino acidsa
Type of equivalent
pseudocenter pairs
Donor
Acceptor
Donor
Acceptor
Donor
Acceptor
Acceptor
PI
Acceptor
Aliphatic
Aliphatic
Aliphatic
Aliphatic
Aliphatic
Donor
Donor
Aliphatic
1ecm (E. coli )
K39
V46
D48
D48
E52
E52
I81
S84
Q88
V35
K39
V46
I81
V85
R11
R11
I14
A
A
A
A
A
A
A
A
A
A
A
A
A
A
B
B
B
4csm (S. cerevisiae )
s
p
p
p
p
s
p
p
p
s
s
s
s
s
s
s
s
K168
I192
N194
N194
E198
E198
I239
T242
D246
V164
K168
I192
I239
K243
R16
R16
L19
B
B
B
B
B
B
B
B
B
B
B
B
B
B
B
B
B
s
p
p
p
p
s
p
p
s
s
s
s
s
s
s
s
s
a
One-letter residue name, residue number, chain-ID and origin (s: center originates from side-chain atom(s); p: center originates
from backbone atom).
398
Detecting Related Function Among Proteins
Figure 8. Equivalent phosphate
binding areas in the binding pockets of uridylate kinase (1ukz) and
the structure of a kinesine-type
domain (3kar). The superposition
based on the matching pseudocenters shows extensive conservation
of the involved amino acids (I) and
the common surface patches (II).
The phosphate groups of the
bound ADP-ligands superimpose
well (III).
common subgraph matches. Finally, the dataset of
5448 cavities has been further clustered into 1010
parent sets from which the query cavities were
selected and subsequently compared against the
remaining set of probe cavities.
Validation of the cavity matching algorithm
In the following we will use a representative set
of bench mark examples, partly taken from literature, to assess and demonstrate the scope, applicability and success rate of our new approach. As a
first test, we retrieve the binding pocket from the
same protein present in different species sharing
low sequence homology. We then move to the
detection of binding sites accommodating the
same ligand. Finally, binding sites are compared
that catalyze similar chemical reactions.
Similarity between binding pockets of two
chorismate mutases originating from
different species
As a first comparative example we selected two
binding pockets extracted from chorismate
mutases originating from two species: Saccharomyces cerevisiae and Escherichia coli.67,68 This
example has previously been studied by Rosen
et al.42 using sparse critical points44 derived from
the Connolly surface of previously extracted binding pockets. In a one-to-one comparison they
could successfully detect a common surface-point
pattern in the two binding pockets. Although the
two proteins show a sequence identity of less than
20% they adopt a similar fold and bind the same
ligand. The bicyclic transition-state-analog inhibitor is recognized in both cases via side-chain interactions. In order to examine whether our
approach is capable to retrieve and match the two
chorismate cavities, we defined the cavity from
S. cerevisiae as query and screened our complete
sample set including the chorismate example from
E. coli. Both scoring criteria R1 and R2 place the
E. coli cavity on the best rank. The actually
obtained match is shown in Figure 7. Table 3 lists
the corresponding pseudocenters with the associated amino acids used in the superpositioning procedure. Although the actual coordinates of the
ligands were not used in the approach, the
obtained cavity surface match generates a trans-
formation that displays the bound inhibitors in a
virtually perfect superposition. Even so the
enzymes show similar fold, ligand recognition
does not necessarily require the same amino acid
composition of the binding site. This result is particularly remarkable since the actual interactions
to the ligands are nearly exclusively performed by
side-chain contacts. Furthermore, in the S. cerevisiae
enzyme the binding pocket is composed by residues emerging from one peptide chain whereas in
the E. coli protein two chains are contributing
(see Figure 6). This fact would clearly limit the
applicability of sequence alignment methods to
detect cavity similarity.
Similarity between portions of co-factor
binding pockets in non-homologous proteins
The successful retrieval of two pockets recognizing the same rigid ligand prompted us to extend
our approach to a larger set of similar and more
flexible ligands. Cofactors are frequently found as
common ligands in proteins, accordingly their
binding has been matter of comparative studies.
Already in 1984, Hol & Wierenga69 detected common binding-site features next to phosphate
groups of bound ligands. Extended a-helical structural motifs generate a partially charged, highly
polar binding region favorably occupied by negatively charged phosphate groups. Kinoshita et al.70
studied the local environment of phosphate groups
in nucleotide-binding proteins via the comparison
of all protein atom coordinates in a sphere of 7 Å.
Obviously structural similarity has been detected.
Usually several NH groups either of the backbone
or side-chains point towards the phosphate binding site. In our approach we selected the local
phosphate recognition site of a kinesine-type
domain (3kar)71 accommodating ADP and queried
the observed pattern against other proteins. The
cavity from a uridylate kinase (1ukz)72 shows high
local similarity giving rise to the mutual alignment
presented in Figure 8. Actually this latter enzyme
exhibits no sequence and fold similarity with the
protein from where the query cavity had been
extracted. Nevertheless, it similarly recognizes the
phosphate position of an ADP.
As another example for local pattern matching,
Kobiyashi et al.38 and Moodie et al.39 investigated
on the basis of well selected data samples the local
399
Detecting Related Function Among Proteins
Figure 9. Pseudocenter and surface patch patterns shared in common between two adenine-binding
regions. The underlying protein
structures show no significant
sequence and fold homology; (I)
depicts the superposition of the
matching pseudocenters of 1cdk
and 1aog (yellow or white carbon
atoms, respectively) and (II) the
superposition of the corresponding
surface patches. In (III), the bound
ligands are also displayed being 50 adenyl– imido-triphosphate (1cdk,
carbons in green) and flavine-adenine-dinucleotide (1aog). In (IV)
and (V) the corresponding interactions with the respective adenine
fragments are illustrated.
binding environment of adenine portions in their
protein receptors. In the first study, a manual binding-site superposition has been performed whereas
Moodie et al. detected a conserved pattern of
physicochemical properties apart from the actual
amino acid composition. In our analysis we
refrained from a sophisticated pre-selection of a
data sample. Instead we picked by chance the
pocket
from
a
cAMP-dependent
protein
kinase (1cdk)73 accommodating 50 -adenyl-imidotriphosphate as ligand. The query pocket has been
edited to display the local environment adjacent to
the adenine portion. To avoid any trivial similarity
matches all entries were removed possessing
high sequence identity with 1cdk according to
the assignment in the PDB-select database. The
remaining 5431 entries have been used for the
mutual comparison with the query pocket.
Based on the R1 scoring, cavities have been
detected on the first ranks that adopt, according
to the FSSP score, the same fold as the reference,
Table 4. Equivalent pseudocenter pairs and the involved
amino acids found in the adenine binding regions of the
structures 1cdk and 1aog used in the cavity matching
algorithm
Corresponding amino acidsa
1cdk
1aog
Type of equivalent
pseudocenter pairs
Acceptor
Acceptor
Donor
Acceptor
PI
Aliphatic
Aliphatic
a
L49
E121
V123
V123
F327
A70
L173
B
B
B
B
B
B
B
p
p
p
p
s
s
s
D36
G128
G128
G126
D36
I11
V37
A
A
A
A
A
A
A
s
p
p
p
p
s
s
One-letter residue name, residue number, chain-ID and
origin (s: center originates from side-chain atom(s); p: center
originates from backbone atom).
however with no significant sequence homology.
Already on rank 7, the cavity extracted from
trypanothione reductase (1aog)74 possessing no
sequence and fold homology with 1cdk, is found
(Figure 9). The matching pseudocenters are listed
in Table 4. This example demonstrates that an
extensive correspondence in surface portions or
pseudocenters does not necessarily result from a
close spatial alignment of the contributing amino
acid residues, but more important from a firm
resemblance of spatial physicochemical properties
in space. Considering the actually bound ligands,
a convincing spatial match of the adenine moiety
in 50 -adenyl-imido-triphosphate is found with that
of flavine-adenine-dinucleotide in 1aog.
Matching entire co-factor binding pockets
In the previous examples only binding subcavities recognizing recurrently similar ligand portions have been investigated. To examine the
scope of our approach, we selected NADPH as a
much larger ligand frequently observed in protein
structures. As query cavity we extracted the pocket
in a carbonyl reductase (1cyd).75 To avoid trivial
matches any proteins with high sequence
homology to 1cyd have been discarded from our
sample set yielding 5377 cavities. As a result from
this search on the first ranks based on R1 or R2
only cavities from other oxidoreductases have
been detected (see Supplementary Material), all
accommodating either a complete NADP(H) or
NAD(H) ligand. Cavities placed at minor ranks
host co-factors with decreasing similarity however
still containing parts of the NADPH skeleton.
Approximately from rank 50 onwards only parts
of the NADPH co-factor binding pocket are recognized, occasionally because the co-factor adopts a
deviating conformation from that in the query
400
Detecting Related Function Among Proteins
Figure 10. The first 300 best-ranked solutions from a comparison of the binding site in the trypsin structure 1tpo
against a set of 5284 probe cavities are shown. Sorting has been performed according to R1 and R2. The first example
for a cavity from the subtilisin-family is found on rank 95 (R1) or on rank 113 (R2).
cavity. Furthermore, at these ranks cavities are
detected hosting other ligands usually originating
from proteins with different function. At rank 98,
the cavity of a methyl transferase (2adm),76 binding
S-adenosylmethionine is found and at rank 116 the
pocket of a phenol hydrolase (1foh)77 is observed
that accommodates FAD as co-factor. Here, the
adenine-recognition site is shared in common with
the original NADPH query pocket.
The present example provides another remarkable insight. On rank 41 the large cavity of a
NADPH-dependent
steroid
dehydrogenase
(1fds)78 is detected, although its crystal structure
has been determined in the absence of the bound
co-factor. Only because a steroid is present in this
large pocket, the entry remained in our data
sample. Nevertheless, the co-factor cavity observed
in the carbonyl reductase matches well with the
large unoccupied part of the pocket in the steroid
dehydrogenase and falls next to the binding site
of the steroid. Thus our approach can also be used
to detect and match unfilled binding sites in a
comparative analysis.
Recognition of binding cavities in proteins of
similar biochemical function
The previous case study convinced us that our
approach should be suited to retrieve proteins of
common biochemical function from an extended
sample set. For our analysis we selected serine proteases. This example has previously been studied.
Fischer et al.41 presented a geometric hashing algorithm to detect similarities among serine proteases.
However, they used molecular descriptors taken
from the entire protein structure for their analysis.
These reflect more strongly features based on the
protein fold rather than our method that reflects
physicochemical properties experienced by the
binding-site residues only.
In a first run, the binding pocket of the trypsin
structure 1tpo79 has been selected as query cavity.
The remaining set has been reduced to 5248 entries
by discarding those cases that were expected to
produce trivial similarity solutions (criteria applied
Table 5. Equivalent pseudocenter pairs and the involved
amino acids found in catalytic centers of two non-equivalent serine proteases from the trypsin (1tpo) and subtilisin-like family (2prk)
Figure 11. Superposition of the cavity from trypsin
(1tpo, white carbon atoms) and from proteinase K (2prk,
yellow carbon atoms) based on the pseudocenter
patterns matched as similar in both cases. Obviously,
from the catalytic triad the histidine, and the serine are
considered together with the oxy-anion hole and the
non-specific peptide recognition site. The catalytic aspartate is not surface-exposed. The superimposed amino
acids (I) and surface patches (II) are shown.
Type of equivalent
pseudocenter pairs
Corresponding amino acidsa
1tpo
2prk
Aromatic
Acceptor
Acceptor
Donor
Donor
Donor/Acceptor
PI
Acceptor
PI
Acceptor
Acceptor
Acceptor
H57
D189
D189
G193
S195
S195
S214
S214
W215
W215
G216
S217
a
s
s
s
p
p
s
p
p
p
p
p
p
H69
A159
A158
N161
S224
S224
S132
S132
L133
L133
G134
G135
s
p
p
p
p
s
p
p
p
p
p
p
One-letter residue name, residue number and origin (s: center originates from side-chain atom(s); p: center originates from
backbone atom).
401
Detecting Related Function Among Proteins
Figure 12. (I) The superposition
of the amino acids of trypsin (1tpo,
carbon atoms colored in yellow)
and a ketosteroid isomerase (1qjg,
carbon atoms in white). The ketosteroid isomerase was found on
rank 120. In (II) the superposition
of surface patches is shown. The
matching pattern comprises several
pseudocenters involved in the peptide backbone exposed to the binding pocket.
as above). The comparison was ranked according
to R1 and R2 (Figure 10) (see Supplementary
Material). On the top ranks only other members of
the trypsin family were detected such as thrombin,
chymotrypsin or tryptase. These are followed by
other examples adopting similar fold, however
exhibiting decreasing sequence similarity (e.g.
kallikrein A, factor D, a-lytic protease or
proteinase A). At rank 113, thus among the first
3% of the data sample considered, a binding
pocket from the structurally unrelated subtilisin
family (1sue)80 has been registered. Trypsin and
subtilisin are both representative parent structures
of the two major serine protease classes. They
share the same biochemical and mechanistic function, however without sequence and fold homology. On the following ranks other examples of
the subtilisin family were found. Comparing the
individual matches of pseudocenters or common
surface patches among members of the two
families shows that the physicochemical properties
of the catalytic serine and histidine are matched
(the aspartate is not surface-exposed) along with
the oxyanion hole and the binding features
experienced by the non-specific peptide recognition site (Figure 11 and Table 5).
Apparently a scoring based on R2 is better discriminating, however based on R1 members of the
subtilisin family are recognized on even more
prominent ranks. A critical assessment of the
results should not ignore that our approach also
produces solutions that appear on a first glance of
no relevance for the detection of common functional features. On rank 120, a common patch is
shown between the query trypsin cavity and that
extracted from a ketosteroid isomerase (1qjg).81
The common pattern expands in this case over several pseudocenters assigned to atoms involved in
the peptide backbone. According to the given
molecular dimensions in such structural elements
a common and obviously rather repetitive pattern
is detected by the clique algorithm once such a
unit is exposed to the cavity surface (Figure 12).
To assess the reliability of our approach we
inverted the “serine-protease” query, now selecting
the subtilisin binding pocket of 1sua82 as query
cavity. In this query other subtilisin cavities were
not excluded. As expected this run retrieves at
first the other entries from the subtilisin family followed by proteinase K and members of trypsintype family on the subsequent ranks. The listing
of entries clearly shows that the data base is less
populated of structural variants by the subtilisin
family (see Supplementary Material).
Idea generator for de novo design
De novo design seeks for novel ligand skeletons
to occupy a given binding pocket. The detection of
common surface patches among binding pockets
might provide some new ideas about possible
lead structures via the analysis of the actual cavity
occupants. However, it should be noted that
besides the initial selection of the cavity data set,
no ligand information is used in the approach.
Interestingly enough we detected a surface patch
of an adenine-binding pocket to be similar with an
unoccupied binding-site region in HIV protease.
In a study by Martin et al.83 the binding of a series
of macrocyclic peptidomimetic inhibitors to HIV
protease is described. Depending on the substitution pattern these ligands orient different
Figure 13. Binding site of the
HIV-protease (1b6o) with a bound
macrocyclic peptidomimetic inhibitor. For reasons of clarity some
amino acids are not shown. An
unoccupied area exhibiting similar
physicochemical properties compared to an adenine-binding site
present in 1cdk is indicated. On the
right, a schematic illustration of the
most important interactions formed
to the inhibitor are plotted.
402
Detecting Related Function Among Proteins
Figure 14. Area of physicochemical equivalence found in the
unoccupied portion of the binding
site in HIV protease 1b6o (Ia) and
the adenine-binding cavity in 1cdk
(IIa). The interactions matched as
equivalent are shown in (Ib) and
(IIb).
molecular portions into the S2’ subpocket, e.g. a
p-aminosulfonamide group. However, in the pdb
entry 1b6o the bound ligand lacks a phenylsulfonyl group at this position, thus leaving the
addressed binding-site niche unoccupied. The
corresponding surface patch of this niche occurs
similarly in the catalytic subunit of protein
kinase A (Figure 13 and Figure 14). There the
patch accommodates the adenine portion of
adenylamino diphosphate. The discovery of such
examples could be of potential interest to de novo
design of protein ligands. Molecular building
blocks detected by this approach are actually
known to be recognized at a site with a particular
protein surface pattern. Possibly they can be joint
with a ligand that occupies the remaining part of
the binding site. The present HIV example stimulated us to perform a more detailed search using
the described binding-site niche as a query cavity.
The search against a dataset of 7192 probe cavities
revealed several examples besides ligands containing an adenine moiety where this patch is occupied
by a hydrophobic aromatic group being part of a
larger ligand, e.g. a p-hydroxybenzamidyl group
in the cAMP-dependent protein kinase inhibitor
balanol (1bx6)84 (Figure 15). This finding matches
well with the fact that the structural studies of
Tyndall et al. actually show in one of the HIV complexes a phenyl group filling up this niche (1d4 l).85
These last examples demonstrate the potential use
of our approach in ligand de novo design. Likely, a
large enough database of cavities together with
their bound ligands can be used to generate interesting suggestions how to modify and improve
known protein ligands.
Conclusions and Outlook
As a basic concept, our new approach assumes
that similar function among proteins requires similar binding pockets. These pockets have to expose
spatially conserved physicochemical properties in
order to recognize and subsequently respond to
Figure 15. Superposition of the
binding sites of HIV-protease 1b6o
(carbon atoms in white) and protein
kinase 1bx6 (carbon atoms in
yellow) is given in (I). The
p-hydroxy-benzamidyl group of
the kinase ligand binds right into
the hydrophobic niche where the
phenyl-sulfonyl group of the 1b6o
ligand is missing (see the text for
further explanation). For comparative purposes the superposition of
HIV-protease 1d4 l with 1bx6 is displayed in (II), showing the close
overlap of the aromatic moiety in
the hydrophobic niche.
403
Detecting Related Function Among Proteins
the binding of the same or a related substrate or
endogenous ligand. Following this idea we depart
from the actual sequence or fold information and
move to a more general description of features
that determine conserved interaction patterns. We
developed a unique coding scheme to condense
the properties of cavity-flanking residues into
simple descriptors. Common recognition patterns
in terms of conserved subsets of these associated
descriptors are detected using a clique detection
algorithm. However, to retrieve relevant information a reliable scoring scheme is essential that
measures surface patches shared among the
matched sub-pockets in common.
Using a representative set of benchmark
examples we could demonstrate the scope of our
approach. It extracts and matches from a sample
set of several thousand cavities extracted from
non-homologous proteins those examples that
recognize the same ligand. Equally well entire
co-factor binding sites can be retrieved. With
decreasing similarity to the skeleton of the reference co-factor, the method also matches only subpockets shared in common with the query cavity
accommodating the reference co-factor. Functional
relationships among proteins resulting in the
catalysis of a similar enzymatic reaction can be
retrieved with our approach. However, in our
opinion the most promising aspect is its potential
to suggest alternative molecular building blocks in
de novo design. The search for putative molecular
portions well-suited to accommodate a particular
sub-pocket of the binding site under consideration
can be inspired by the retrieval of ligands actually
occupying a very similar sub-pocket in other
already structurally characterized proteins. Such a
source of information for ligand design may
develop a routine tool in supporting structurebased drug design.
A further aspect could be of potential relevance
in understanding drug action. Frequently, side
effects of drugs are created due to undesired binding to the pocket of another protein. A search
based on the described approach provides the
possibility to detect structurally related binding
cavities where such unexpected binding could
occur. Important enough this prediction does not
rely on the ligand properties but purely on the
shape of the binding pockets. This leaves room to
modify a ligand structurally to achieve better
selectivity.
The present approach requires several improvements. First-of-all the classification of exposed
amino acids has to cover all relevant interaction
patterns. At present our approach neglects some
supposedly important contact geometries (e.g.
p-stacking to carboxy or guanidine groups).
Secondly, the scoring of the different solutions
suggested by the clique detection algorithm is
essential for the retrieval of relevant information.
At present it is based on surface patches shared in
common by the matching cavities. Improved
figures-of-merit have to consider better the con-
tiguous connection of matched surface patches.
Finally, at present our approach is computationally
rather intensive. It requires algorithmic accelerations. Such improvements would enable an allagainst-all comparison of the entire cavity database. Such a study is likely to provide an entirely
new classification and clustering of protein structures in terms of cavity similarity aspects.
Acknowledgements
The authors acknowledge stimulating discussion
with Dr M. Hendlich (Lion Biosciences,
Heidelberg, Germany) in particular in the beginning of this project. The help of Judith Günther
(Univ. Marburg) and Dr A. Bergner (CCDC,
Cambridge, UK) in implementing various aspects
of Cavbase is gratefully acknowledged. We thank
Professor A. Ultsch and R. Simon (University Marburg) for helpful discussions about algorithmic
aspects. The present project has been supported
by the German Minister of Science and
Education (bmb þ f) in the framework of the
ReLiMo project (Grant No. 0311619). We thank all
partners in this project for a fruitful and successful
collaboration.
References
1. Rubin, G. M., Yandell, M. D., Wortman, J. R., Miklos,
G. L. G., Nelson, C. R., Hariharan, I. K. et al. (2000).
Comparative Genomics of the Eukaryotes. Science,
287, 2204– 2215.
2. Broder, S. & Venter, J. C. (2000). Sequencing the entire
genomes of free-living organisms: the foundation of
pharmacology in the new millennium. Annu. Rev.
Pharmacol. Toxicol., 40, 97 –132.
3. Lander, E. S., Linton, L. M., Birren, B., Nusbaum, C.,
Zody, M. C., Baldwin, J. et al. (2001). Initial sequencing and analysis of the human genome. Nature,
409, 860– 921.
4. Marcotte, E. M., Pellegrini, M., Ng, H., Rice, D. W.,
Yeates, T. O. & Eisenberg, D. (1999). Detecting protein function and protein–protein interactions from
genome sequences. Science, 285, 751– 753.
5. Lottspeich, F. (1999). Proteome analysis: a pathway
to the functional analysis of proteins. Angew. Chem.
Int. Ed. 38, 2476–2492.
6. Wang, J. H. & Hewick, R. M. (1999). Proteomics in
drug discovery. Drug. Discov. Today, 4, 129– 383.
7. Danchin, A. (1999). From protein sequence to function. Curr. Opin. Struct. Biol. 9, 363– 367.
8. Orengo, C. A., Todd, A. E. & Thornton, J. M. (1999).
From protein structure to function. Curr. Opin. Struct.
Biol. 9, 374– 382.
9. Westhead, D. R. & Thornton, J. M. (1998). Protein
structure prediction. Curr. Opin. Biotechnol. 9,
383 –389.
10. Blundell, T., Jhoti, H. & Abell, C. (2002). Highthroughput crystallography for lead discovery in
drug design. Nature Rev. Drug Discov. 1, 45 – 54.
11. Burley, S. K., Almo, S. C., Bonanno, J. B., Capel, M.,
Chance, M. R., Gaasterland, T., Lin, D., Sali, A.,
404
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
27.
28.
29.
30.
Studier, F. W. & Swaminathan, S. (1999). Structural
genomics: beyond the Human Genome Project.
Nature Genet. 23, 151– 157.
Rost, B. (1998). Marrying structure and genomics.
Structure, 6, 259– 263.
Thornton, J. M., Todd, A. E., Milburn, D., Borkakoti,
N. & Orengo, C. A. (2000). From structure to function: approaches and limitations. Nature Struct. Biol.
7, 991– 994.
Laskowski, R. A., Luscombe, N. M., Swindells, M. B.
& Thornton, J. M. (1996). Protein clefts in molecular
recognition and function. Protein Sci. 5, 2438–2452.
Needleman, S. B. & Wunsch, C. D. (1970). A general
method applicable to the search for similarities in
the amino acid sequence of two proteins. J. Mol.
Biol. 48, 443– 453.
Waterman, M. S. (1984). General methods for
sequence comparison. Bull. Math. Biol. 46, 473–500.
Taylor, W. R. & Orengo, C. A. (1989). Protein structure alignment. J. Mol. Biol. 208, 1 – 22.
Pearson, W. R. & Lipman, D. J. (1988). Improved
Tools for Biological Sequence Analysis. Proc. Natl
Acad. Sci. USA, 85, 2444– 2448.
Pearson, W. R. & Lipman, D. J. (1990). Rapid and
Sensitive Sequence Comparison with FASTP and
FASTA. Methods Enzymol., 183, 63 – 98.
Bleasby, A. J., Akrigg, D. & Attwood, T. K. (1994).
OWL: A non-redundant, composite protein sequence
database. Nucl. Acids Res. 22, 3574– 3577.
Bairoch, A. & Boeckmann, B. (1994). The S.W.I.S.SP.R.O.T. protein sequence databank: current status.
Nucl. Acids Res. 22, 3578– 3580.
Lessel, U. & Schomburg, D. (1994). Similarities
between protein 3D structures. Protein Eng. 7,
1175 –1187.
Holm, L. & Sander, C. (1993). Protein structure comparison by alignment of distance matrices. J. Mol.
Biol. 233, 123– 138.
Grindley, H. M., Artymiuk, P. J., Rice, D. W. &
Willett, P. (1993). Identification of tertiary structure
resemblance in proteins using a maximal common
subgraph isomorphism algorithm. J. Mol. Biol. 229,
707– 721.
Nussinov, R. & Wolfson, H. J. (1991). Efficient detection of three-dimensional structural motifs in biological macromolecules by computer vision
techniques. Proc. Natl Acad. Sci. USA, 88,
10495– 10499.
May, A. C. W. & Johnson, M. S. (1995). Improved
genetic algorithm-based protein structure comparison. Protein Eng. 8, 873– 882.
May, A. C. W. & Johnson, M. S. (1994). Protein structure comparison using a combination of a genetic
algorithm, dynamic programming and least-squares
minimization. Protein Eng. 7, 475– 485.
Lehtonen, J. V., Dennessiouk, K., May, A. C. W. &
Johnson, M. S. (1999). Finding local structural
similarities among families of unrelated protein
structures: a generic non-linear alignment algorithm.
Proteins: Struct. Funct. Genet., 34, 341– 355.
Richards, F. M. & Kundrot, C. E. (1988). Identification
of structural motifs from protein coordinate data:
secondary structure and first-level supersecondary
structure. Protein: Struct. Funct. Genet. 3, 71 – 84.
Mitchell, E. M., Artymiuk, P. J., Rice, D. W. & Willett,
P. (1989). Use of graph theory to compare secondary
structure motifs in proteins. J. Mol. Biol. 212,
151– 166.
Detecting Related Function Among Proteins
31. Sali, A. & Blundell, T. L. (1990). Definition of general
topological equivalence in protein structures. A
procedure involving comparison of properties and
relationships through simulated annealing and
dynamic programming. J. Mol. Biol. 4, 403– 428.
32. Vriend, G. & Sander, C. (1991). Detection of common
three-dimensional substructures in proteins. Proteins:
Struct. Funct. Genet., 11, 52 – 58.
33. Alexandrov, N. N., Takahashi, K. & Go, N. (1992).
Common spatial arrangements of backbone fragments in homologous and non-homologous proteins.
J. Mol. Biol. 225, 5– 9.
34. Pennec, X. & Ayache, N. (1998). A geometric algorithm to find small but highly similar 3D substructures in proteins. Bioinformatics, 14, 516– 522.
35. Wallace, A. C., Borkakoti, N. & Thornton, J. M.
(1997). TESS: A geometric hashing algorithm for
deriving 3D coordinate templates for searching
structural databases. Application to enzyme active
sites. Protein Sci. 6, 2308– 2323.
36. Artymiuk, P. J., Poirrette, A. R., Grindley, H. M., Rice,
D. W. & Willett, P. (1994). A graph-theoretic approach
to the identification of three-dimensional patterns of
amino acid side-chains in protein structures. J. Mol.
Biol. 243, 327– 344.
37. Russell, R. B. (1998). Detection of protein threedimensional side-chain patterns: new examples of
convergent evolution. J. Mol. Biol. 279, 1211 – 1227.
38. Kobayashi, N. & Go, N. (1997). A method to search
for similar protein local structures at ligand-binding
sites and its application to adenin recognition. Eur.
Biophys. J. 26, 135– 144.
39. Moodie, S. L., Mitchell, J. B. & Thornton, J. M. (1996).
Protein recognition of adenylate: an example of a
fuzzy recognition template. J. Mol. Biol. 263, 486– 500.
40. Stahl, M., Taroni, C. & Schneider, G. (2000). Mapping
protein surface cavities and prediction of enzyme
class by a self-organizing neuronal network. Protein
Eng. 13, 83 – 88.
41. Fischer, D., Wolfson, H., Lin, S. L. & Nussinov, R.
(1994). Three-dimensional, sequence order-independent structural comparison of a serine protease
against the crystallographic database reveals active
site similarities: potential implications to evolution
and to protein folding. Protein Sci. 3, 769– 778.
42. Rosen, M., Liang, S. L., Wolfson, H. & Nussinov, R.
(1998). Molecular shape comparisons in searches for
active sites and functional similarity. J. Mol. Biol. 11,
263– 277.
43. Laskowski, R. A. (1995). SURFNET: A program for
visualizing surfaces, cavities and intermolecular
interactions. J. Mol. Graph. 31, 2735– 2748.
44. Lin, S. L., Nussinov, R., Fischer, D. & Wolfson, H. J.
(1994). Molecular surface representations by sparse
critical points. Proteins: Struct. Funct. Genet., 18,
94 –101.
45. Klebe, G. (1994). The use of composite crystal-field
environments in molecular recognition and the de
novo design of protein ligands. J. Mol. Biol. 237,
212– 235.
46. Klebe, G. (1993). Structural alignment of molecules.
In 3D QSAR and Drug Design: Theory, Methods and
Applications (Kubinyi, H., ed.), pp. 173– 199,
ESCOM, Leiden.
47. Hemm, K., Aberer, K. & Hendlich, M. (1995). Constituting a receptor-ligand information base from
quality-enriched data. Ismb. 3, 170– 178.
405
Detecting Related Function Among Proteins
48. Hendlich, M. (1998). Databases for protein-ligand
complexes. Acta Crystallogr. D Biol. Crystallog. 54
(Pt 6), 1178– 1182.
49. Kuntz, I. D., Blaney, J. M., Oatley, S. J., Langridge, R.
& Ferrin, T. E. (1982). A geometric approach to
macromolecule-ligand interactions. J. Mol. Biol. 161,
269–288.
50. Levitt, D. G. & Banaszak, L. J. (1992). POCKET: A
computer graphics method for identifying and displaying protein cavities and their surrounding
amino acids. J. Mol. Graph. 10, 229– 234.
51. Hendlich, M., Rippmann, F. & Barnickel, G. (1997).
LIGSITE: Automatic and efficient detection of potential small molecule-binding sites in proteins. J. Mol.
Graph. Model. 15, 359– 363.
52. Brady, G. P. J. & Stouten, P. F. (2000). Fast prediction
and visualization of protein binding pockets with
PASS. J. Comput. Aided Mol. Des. 14, 383– 401.
53. Liang, J., Edelsbrunner, H. & Woodward, C. (1998).
Anatomy of protein pockets and cavities: measurement of binding site geometry and implications for
ligand design. Protein Sci. 7, 1884– 1897.
54. Edelsbrunner, H. & Mücke, E. P. (1994). Threedimensional alpha shapes. ACM Trans. Graph. 13,
43– 72.
55. Edelsbrunner, H. (1995). The union of balls and its
dual shape. Discrete Comput. Geom. 13, 415– 440.
56. Peters, K. P., Fauck, J. & Frömmel, C. (1996). The
automatic search for ligand binding sites in proteins
of known three-dimensional structure using only
geometric criteria. J. Mol. Biol. 256, 201– 213.
57. Hendlich, M., Bergner, A., Günther, J. & Klebe, G.
(2002). Relibase—design and development of a database for comprehensive analysis of protein-ligand
interactions. J. Mol. Biol., 00 submitted.
58. Günther, J., Bergner, A., Hendlich, M. & Klebe, G.
(2002). Utilising structural knowledge in drug design
strategies—applications using Relibase. J. Mol. Biol.,
00 submitted.
59. Bruno, I. J., Cole, J. C., Lommerse, J. P., Rowland,
R. S., Taylor, R. & Verdonk, M. L. (1997). IsoStar: A
library of information about nonbonded interactions.
J. Comput.-Aided Mol. Des., 11, 525– 537.
60. Bron, C. & Kerbosch, J. (1973). Algorithm 457. Finding all cliques of an undirected graph. Commun.
ACM, 16, 575–577.
61. Brint, A. T. & Willett, P. (1987). Algorithms for
the identification of three-dimensional maximal
common substructures. J. Chem. Inf. Comput. Sci. 27,
152–158.
62. Brint, A. T. & Willett, P. (1989). Upperbound procedures for the identification of similar three-dimensional chemical structures. J. Comput.-Aided Mol.
Design, 2, 311 –320.
63. Sayle, R. A. & Milner-White, E. J. (1995). RASMOL:
biomolecular graphics for all. Trends Biochem. Sci. 20,
374.
64. Fischer, D., Tsai, C. J., Nussinov, R. & Wolfson, H.
(1995). A 3D sequence-independent representation
of the protein data bank. Protein Eng, 8, 981– 997.
65. Hobom, U., Scharf, M., Schneider, R. & Sander, C.
(1992). Selection of a representative set of structures
from the Brookhaven Protein Data Bank. Protein Sci.
1, 409–417.
66. Hobom, U. & Sander, C. (1994). Enlarged representative set of protein structures. Protein Sci. 3, 522.
67. Strater, N., Schnappauf, G., Braus, G. & Lipscomb,
W. N. (1997). Mechanisms of catalysis and allosteric
68.
69.
70.
71.
72.
73.
74.
75.
76.
77.
78.
79.
80.
regulation of yeast chorismate mutase from crystal
structures. Structure, 5, 1437– 1452.
Lee, A. Y., Karplus, P. A., Ganem, B. & Clardy, J.
(1995). Atomic structure of the buried catalytic
pocket of Escherichia coli chorismate mutase. J. Am.
Chem. Soc. 117, 3327– 3362.
Hol, W. & Wierenga, R. K. (1984). The a-helix dipole
and the binding of phosphate groups of coenzymes
and substrates by proteins. In X-ray Crystallography
and Drug Design (Horn, A. S. a. D. R.,C. J., ed.), ???,
Oxford.
Kinoshita, K., Sadanami, K., Kidera, A. & Go, N.
(1999). Structural motif of phosphate-binding site
common to various protein superfamilies: Allagainst-all structural comparison of protein-mononucleotide complexes. Protein Eng. 12, 11 – 14.
Gulick, A. M., Song, H., Endow, S. A. & Rayment, I.
(1998). X-ray crystal structure of the yeast Kar3
motor domain complexed with Mg.ADP to 2.3 Å
resolution. Biochemistry, 37, 1769– 1776.
Muller-Dieckmann, H. J. & Schulz, G. E. (1995). Substrate specificity and assembly of the catalytic center
derived from two structures of ligated uridylate
kinase. J. Mol. Biol. 246, 522– 530.
Bossemeyer, D., Engh, R. A., Kinzel, V., Ponstingl, H.
& Huber, R. (1993). Phosphotransferase and substrate binding mechanism of the cAMP-dependent
protein kinase catalytic subunit from porcine heart
as deduced from the 2.0 Å structure of the complex
with Mn2 þ adenylyl imidodiphosphate and inhibitor peptide PKI(5-24). EMBO J. 12, 849–859.
Zhang, Y., Bond, C. S., Bailey, S., Cunningham, M. L.,
Fairlamb, A. H. & Hunter, W. N. (1996). The crystal
structure of trypanothione reductase from the
human pathogen Trypanosoma cruzi at 2.3 Å resolution. Protein Sci. 5, 52 – 61.
Tanaka, N., Nonaka, T., Nakanishi, M., Deyashiki, Y.,
Hara, A. & Mitsui, Y. (1996). Crystal structure of the
ternary complex of mouse lung carbonyl reductase
at 1.8 Å resolution: the structural origin of coenzyme
specificity in the short-chain dehydrogenase/
reductase family. Structure, 4, 33 – 45.
Schluckebier, G., Kozak, M., Bleimling, N., Weinhold,
E. & Saenger, W. (1997). Differential binding of
S-adenosylmethionine S-adenosylhomocysteine and
Sinefungin to the adenine-specific DNA methyltransferase M.TaqI. J. Mol. Biol. 265, 56 – 67.
Enroth, C., Neujahr, H., Schneider, G. & Lindqvist, Y.
(1998). The crystal structure of phenol hydroxylase in
complex with FAD and phenol provides evidence for
a concerted conformational change in the enzyme
and its cofactor during catalysis. Structure, 6,
605 –617.
Breton, R., Housset, D., Mazza, C. & FontecillaCamps, J. C. (1996). The structure of a complex of
human 17b-hydroxysteroid dehydrogenase with
estradiol and NADP þ identifies two principal
targets for the design of inhibitors. Structure, 4,
905 –915.
Marquart, M., Walter, J., Deisenhofer, J., Bode, W. &
Huber, R. (1983). The geometry of the reactive site
and the peptide groups in trypsin, trypsinogen and
its complexes with inhibitors. Acta Crystallog. B39,
480 –490.
Gallagher, D. T., Pan, Q. W. & Gilliland, G. L. (1998).
Mechanism of ionic strength dependence of crystal
growth rates in a subtilisin variant. J. Cryst. Growth,
193, 665– 673.
406
81. Cho, H. S., Ha, N. C., Choi, G., Kim, H. J., Lee, D.,
Oh, K. S. et al. (1999). Crystal structure of delta(5)-3ketosteroid isomerase from Pseudomonas testosteroni
in complex with equilenin settles the correct
hydrogen bonding scheme for transition state
stabilization. J. Biol. Chem. 274, 32863– 32868.
82. Almog, O., Gallagher, T., Tordova, M., Hoskins, J.,
Bryan, P. & Gilliland, G. L. (1998). Crystal structure
of calcium-independent subtilisin BPN’ with
restored thermal stability folded without the prodomain. Proteins: Struct. Funct. Genet., 31, 21 – 32.
83. Martin, J. L., Begun, J., Schindeler, A.,
Wickramasinghe, W. A., Alewood, D., Alewood, P. F.
et al. (1999). Molecular recognition of macrocyclic
peptidomimetic inhibitors by HIV-1 protease. Biochemistry, 38, 7978– 7988.
84. Narayana, N., Diller, T. C., Koide, K., Bunnage, M. E.,
Nicolaou, K. C., Brunton, L. L. et al. (1999). Crystal
structure of the potent natural product inhibitor
balanol in complex with the catalytic subunit of
cAMP-dependent protein kinase. Biochemistry, 38,
2367– 2376.
85. Tyndall, J. D., Reid, R. C., Tyssen, D. P., Jardine, D. K.,
Todd, B., Passmore, M. et al. (2000). Synthesis, stability, antiviral activity, and protease-bound structures
of substrate-mimicking constrained macrocyclic
Detecting Related Function Among Proteins
inhibitors of HIV-1 protease. J. Med. Chem., 43,
3495– 3504.
Edited by R Huber
(Received 20 March 2002; received in revised form 10
July 2002; accepted 17 July 2002)
http://www.academicpress.com/jmb
Supplementary Material for this paper comprising three tables listing the best ranked solutions
for the comparisons of a NADPH binding pocket
(1cyd), a trypsin cavity (1tpo) and a subtilisin cavity (1sua) are available on IDEAL